fall detection algorithm based on triaxial accelerometer ... · fall detection algorithm based on...

Fall Detection Algorithm Based on TriaxialAccelerometer and MagnetometerTianjiao Shi, Xingming Sun, Zhihua Xia, Leiyue Chen, and Jianxiao Liu

Abstract—Fall is a precipitous drop from a height, or from ahigher position, which may be accompanied by injuries. This isone of the most dangerous and fearful situation in the elderlyliving. This is the reason, fast and early detection of the fallis very important to save and rescue the people and avoid thebadly prognosis. In this article we are presenting a threshold-based fall detection algorithm that processes data from commonsensors in modern smart phones, such as triaxial accelerometerand magnetometer in order to detect falls. The algorithm usesSignal Vector Magnitude (SVM) peak value, base length andpost-impact velocity to distinguish falls from most of dailyactivities. However, the SVM curve in a period produced byrunning is similar to a fall (a running curve can be regarded asthe combination of multiple fall curves). Accordingly, residualmovement is taken into account to identify running and fall.In addition, the vertical acceleration is observed to increasedetection accuracy. In the experiments, the data are collectedby simulating fall in four directions: forward, backward, leftand right. The simulations are conducted by young people. Thefinal experiment includes data from 120 simulated falls and 150daily activities. Compared with previous methods, the proposedmethod achieves higher sensitivity and specificity.

Index Terms—Accelerometer, Fall Detection, Magnetometer,Residual Movement, Smart phone

I. INTRODUCTION

FALL is a fearful and dangerous in the elderly people’sdaily life. Each year, there are more than one third of

people aged over 65 years old falling [1], [2]. The immediaterescue to fall would greatly alleviate the hurt. The automaticdetection of fall events could help reducing the response timeand significantly improve the prognosis of fall victims.

Nowadays, a number of fall detection methods have beenproposed [3]. According to the methods of data acquirement,fall detection methods can be divided into three categories.The first category can be called context-aware method whichusually utilizes pressure sensor to detect the people whetherfall or not [4], [5]. This kind of method will be out ofwork if the pressure is not under the user’s weight afterfall. The second category is camera-based method which isinstalled as an important part of in-home auxiliary system[6], [7], [8], [9], [10], [11], [12], and the fall events are

Manuscript received June 11, 2015; revised October 06, 2015. This workis supported by the NSFC (61173141, 61232016, 61572258, 61502242,U1405254, U1536206, 61173136, 61373133), 201301030, 2013DFG12860,BK20150925, Fund of Jiangsu Engineering Center of Network Monitor-ing (KJR1308, KJR1402), Fund of MOE Internet Innovation Platform(KJRP1403), CICAEET, and PAPD fund. Zhihua Xia is supported byJiangsu Government Scholarship for Overseas Studies (JS-2014-332).

Tianjiao Shi, Xingming Sun, Zhihua Xia and Leiyue Chen are withJiangsu Engineering Center of Network Monitoring, Jiangsu CollaborativeInnovation Center on Atmospheric Environment and Equipment Technology,School of Computer and Software, Nanjing University of InformationScience & Technology, Nanjing, China. (e-mail: tianjiao [email protected],[email protected], xia [email protected], chenly [email protected]).

Jianxiao Liu is with College of informatics, Huazhong AgriculturalUniversity, China (e-mail: [email protected]).

detected by analysing the video pictures [13], [14], [15], [16],[17], [18], [19] with classification techniques. The wholesystem is usually expensive and only works at the video-surveillance place. The third category is the acceleration-based method which is now the most popularly used one [20],[21], [22], [23], [24], [25], [26], [27], [28], [29], [30]. At thebeginning, the acceleration-based system uses body-attachedsensors such as accelerometers and gyroscopes to acquirekinetic data from human motion. Through it can detect thefall event, it is not portable. The modern smart phoneshave been equipped with various build-in sensors which canbe directly utilized for fall detection [25], [26], [27], [28],[29], [30]. In this paper, we proposed a threshold-based falldetection method by using the triaxial accelerometer andmagnetometer which are common sensors in most kindsof modern smart phones. Four features are extracted fromaccelerometer and magnetometer raw data to distinguish fallevent from activities of daily living (ADL) such as walking,going upstairs, going downstairs, and jumping. In addition,the algorithm also takes residual movement into account afterthe fall to improve the accuracy as a running curve canbe regarded as the combination of multiple fall curves. Aslong as all feature values meet the condition of the thresholdvalues we set, we can recognize the fall.

The remainder of this paper is structured as follows. Insection II, we present an overview of related works. Weintroduce the algorithm design in Section III and imple-mentation in Section IV. In Section V, we evaluate oursystem with extensive experiments. Section VI summarizesthe conclusion of the paper.

II. RELATED WORKS

As mentioned above, the existing fall detection methodscan be divided into three categories. As the development ofsmart phones, the acceleration-based fall detection systemscan be conveniently applied in real world applications, andthus attract more and more attentions.

At the beginning, researchers proposed some acceleration-based fall detection methods by attach the acceleration sensoron tester’s body. It is quite inconvenience, but providespreliminary knowledge for us to refer to. Yang et al. [20]used the two wireless acceleration sensors embedded withNaive Bayes algorithm to implement a wearable real-time falldetection system which has an advantage both in accuracyperformance and model building time. The system attachedsensors to the chest and the thigh provides acceleration infor-mation to detect forward, backward, leftward and rightwardfalls. Wang et al. [24] proposed a low-power fall detectionalgorithm based on triaxial accelerometer and barometricpressure signals. The algorithm dynamically adjusts the sam-pling rate of an accelerometer and manages data transmission

Engineering Letters, 24:2, EL_24_2_06

(Advance online publication: 18 May 2016)

______________________________________________________________________________________

between sensors and a controller to reduce power consump-tion. Pham et al. [23] presented a fall alert system using anaccelerometer sensor for elderly people based on thresholdalgorithm. The algorithm calculates Signal Magnitude Area(SMA) to distinguish fall events from daily activities. If thisvalue is greater than the predefined threshold, the systemrecognizes as a fall and then displays a warning.

For the reason that the update of Micro Electro Mechan-ical System (MEMS) motion sensors technology and thepopularity of mobile phones which are equipped with 3-axis accelerometers and other motion sensors, there is nodoubt that smart phones have become ideal devices for falldetection which are popular in researchers in recent years.

Dai et al. [28] designed a fall detection system, calledPerFall, which is the first pervasive fall detection system us-ing mobile phones. The algorithm is based on accelerometerthresholds: the total acceleration and the vertical accelerationneed to compare with predefined thresholds which is adjustedwith collected data. Sposaro et al. [30] proposed a fall alertsystem called iFall which used triaxial accelerometers for falldetection. The algorithm is based on acceleration magnitudethresholds and timeouts detection. If a fall is detected, thesystem will send a request for help to the caregivers. Tiwariet al. [25] presented a detection mechanism for free fallcondition using an Android based smart phone with anin-built triaxial accelerometer. In this algorithm value ofacceleration is analysed for the time it was close to 0. Inthis time we obtain the net displacement. If the obtain valueof displacement crosses a threshold predefined value thanfree fall condition is confirmed.

III. ALGORITHM DESIGNIn this section, we propose a threshold-based fall detection

algorithm which mainly utilizes four features which areidentified from the analysis of the signal vector magnitude(SVM) to recognize the fall event from human activities.The four features are the SVM peak value, the base lengthof the triangle, post-impact velocity and residual movementrespectively. In addition, the vertical acceleration is addedto the approach as a verified feature. The threshold valuesare chosen based on the data set including different types offall data (including forwards, backwards, sideways) and ADL(including walking, running, jumping, going downstairs andgoing upstairs), as is shown in Fig. 1(c) and Fig. 2.

A. Algorithm Features and ThresholdsHuman body movement can be reflected by acceleration.

When people is walking, going downstairs, going upstairs,jumping, running, or falling, acceleration data varies. Forsimplification, SVM is calculated as follow to eliminatethe sensitivity of mobile phone in aspect of position anddirection.

SVM =√A2

x +A2y +A2

z (1)

where Ax, Ay , and Az represent the accelerating speedsof the X, Y, and Z axes of the accelerometer respectively.SVM values are shown in Fig. 1, the horizontal axis rep-resents the time (ms) and the vertical axis represents theSVM values(m/s2).The algorithm defines the three featuresmentioned in [31] to distinguish fall event from ADL below:

(a) Vertical acceleration

(b) Fall event

(c) Running activity

Fig. 1: This figure shows the waveforms of (a) vertical acceleration, (b)fall event, and (c) running activity. The horizontal axis represents the time

(ms) and the vertical axis represents the SVM values (m/s2).

1) SVM peak value (denoted as P): Most falls have ahigher SVM value than daily activities as shown in Fig.1(b), so the SVM peak value is a good choice in the designof the fall detection algorithm. From the Fig. 2(a), (c), and(d), we can know that the waveforms of walking, goingdownstairs and going upstairs and find that SVM peakvalues of them are all lower than a given thresholds. So,it can detect falls totally but there are still some ADLmistaken for fall.

2) The base length of the triangle (denoted as B): Itis formed by the peak value and the 10 (m/s2) horizontalaxis as shown in the short solid line of Fig. 1(b). When a



______________________________________________________________________________________

fall happens, there is a short time impact and the shock isquickly absorbed by the human body. Therefore, the baselength should be smaller in case of a fall compared to thecase of daily activities.

3) Post-impact Velocity (denoted as V): As what is shownin the shaded area of Fig. 1(b). When the event meets P andB threshold, it can be deemed to a possible fall. The velocityafter the impact is utilized to distinguish sudden moves, likejumping, quickly sitting on a bed or physical exercise. Bythe analysis of the experiment data, P, B, and V values whichproduced the optimal results in terms of accuracy are selectedin Table I:

TABLE I: Fall detection algorithm features and associated optimalthresholds

No Features description Code Threshold Fall sign

1 Peak value for SVM P 23.4 (m/s2) ≥2 Base length B 0.5 s <

3 Post-impact velocity V 0.5 m/s <

Although the three features mentioned above can be usedto separate fall event from most of ADL, running activitycannot be distinguished. So, the algorithm takes residualmovement into account to recognize the running activity.

4) Residual movement: When the fall happens, the firstpeak value reaches the threshold we set. And then, thesecond peak value follows which lower than the thresholdwe set. At last, the acceleration back to normal. However,when we are running, the peak values all exceed the pre-setthreshold and the time interval between the peak values issmaller than the one in fall experiment. The waveforms ofrunning activity and fall event are shown in Fig. 1(b) and(c). Furthermore, we set a variable count in our algorithm,which increases every time in condition that four featuresmentioned above are satisfied in order to describe theresidual movement [32]. If the values of count are between10 and 15, algorithm takes the event as the fall. Otherwise,algorithm takes the event as running.

5) Vertical acceleration: Additionally, fall detection algo-rithm adds the vertical acceleration as a verified procedure.Before and after the fall, the waveform of the vertical acceler-ation is similar to SVM’s. In this study, we utilize the verticalacceleration value |Av| [33] as a verified procedure. Andthe vertical acceleration can also take running-to-walking asnon-fall to improve the accuracy. The vertical accelerationvalue |Av| which is computed as:

|Av| = |Ax sin θz +Ay sin θy −Az cos θy cos θz| (2)

where θy and θz denote its pitch and roll values which isdetermined by a mobile phone’s accelerometer and mag-netometer sensor. If the value of |Av| is larger than thethreshold Th|Av| we pre-set, it will be considered as apossible fall. In the algorithm, we set Th|Av| value for 30.Before and after a fall event, the vertical acceleration valuewill change which is illustrated in Fig. 1(a).

(a) Walking

(b) Jumping

(c) Going downstairs

(d) Going upstairs

Fig. 2: Activities of daily living.



______________________________________________________________________________________

IV. IMPLEMENTATION

We have developed fall detection system which is imple-mented in the smart phone using the algorithm proposed inthis paper. The system mainly consists of three main screensas shown in Fig. 3(a), (b), and (c). Our application is simple,smart and efficient as its biggest advantage. Fig. 3(a) showsthe main screen of the system and the raw data of sensors.Fig. 3(b) depicts some parameters of fall detection algorithmincluding SVM threshold, vertical acceleration threshold andsensor speed value, which can be set by the participant. Fig.3(c) depicts some parameters of warning message includingthe telephone number of the emergency contact, time-outand SMS content which can be set by the participant beforemonitoring application. Once a fall condition is detected, thesystem will send a warning message. The message containsthe participant’s longitude, latitude, detailed address and fallinformation to emergency contact as shown in Fig. 4.

V. EXPERIMENTS

We evaluate the effectiveness of the proposed algorithmwith extensive experiments. In this section, we briefly intro-duce our experimental environment and restricted condition.Then we introduce how the data are collected. In addition,we also present performance of system and compare it withthat of existing algorithms. Finally, we present resourceconsumption of the system.

A. Experimental Environment

In our experiments, we used the smart phone built-inthree-axis accelerometer and magnetometer to realize thefall detection algorithm. The sampling frequencies of theaccelerometer and magnetometer values are both 15Hz. Fur-thermore, results indicate that fall detection using a triaxialaccelerometer attached to the waist or head is more efficient[34] from previous studies. So, all subjects are required toattach the smart phone to the waist to get data compared toattach to the head which is more convenient in our study.

B. Data Collection

Since elderly people are not suitable for simulating falls,our data is derived from healthy young subjects who volun-teer to perform the simulate falls and the ADL tasks.

We selected 10 young subjects, including 6 male and 4female, with an average age of 25 ± 3 years, average heightof 170 ± 10 cm and an average weight of 70 ± 12 kg. Eachparticipant performed four types of falls (including forwards,backwards, sideways). The simulation of each kind of fallwas conducted for 2 minutes and was repeated three times. Inaddition, each participant performed five types of ADL tasks(including walking, running, jumping, going downstairs andgoing upstairs) and each kind of ADL was conducted for5 minutes and was repeated three times. In total, we obtaindata for 120 falls and 150 ADL.

C. Detection Performance

We measure the detection performance in terms of sensi-tivity and specificity. The sensitivity defined in (3) is used tomeasure the algorithm’s ability to determine real falls fromthe set of simulated falls.

Sensitivity =TP

TP + FN×100% (3)

where TP is the true positive representing a fall occurs andthe system properly detects it. FN is the false negative rep-resent a fall occurs but the system does not announce it.

On the other hand, the specificity defined in (4) is used tomeasure the algorithm’s ability to mistake normal activitiesfor falls.

Specificity =TN

TN + FP× 100% (4)

where TN is the true negative representing a non-fall move-ment is performed and the system does not announce a fall.FP is the false positive represent the system announces a fallalthough it did not occur.

• Thresholds RobustnessTo evaluate how robust these thresholds are, we can vary

these thresholds and the algorithm’s performance which canbe re-evaluated. By the analysis of the experiment data, thethresholds listed in Table I produce optimal results. Modify-ing these thresholds can degrade performance. However, itis important to see how much the algorithm’s performancewill change.

First, P, B and V values were changed so that the risk offalse positive or true negative could increase. For instance,if the peak value P becomes smaller, more daily activitiesmay be identified as falls and false positives will increase.Similarly, if the peak value P becomes greater, more fallsmay be identified as daily activities and true negative willincrease.

When the thresholds have a 5% variation to increase thefalse positive, two jumping events and two running eventswere identified as falls. This led to a specificity of 92.67%.When the thresholds have a 10% variation for increasing,false positive was generated by the algorithm, thus obtaining88.67% specificity.

The thresholds were also modified so as to increasethe true negative. When a 5% variation for reducing thethreshold, three falls were misinterpreted as daily activities,which led to 88.83% sensitivity. Similarly, When a 10%variation for reducing the threshold, four fall events weren’tdetected, which led to 87.5% sensitivity. Table II shows theresults obtained using the thresholds variations presented. Ingeneral, the algorithm provided satisfactory results.

TABLE II: Algorithm performance using different threshold values

Threshold valuesAlgorithm performance (%)Sensitivity Specificity

Optimal values(see Table I) 90.83 95.33

5% variation from optimal values so asto increase the risk of false positive

90.83 92.67

10% variation from optimal values soas to increase the risk of false positive

90.83 88.67

5% variation from optimal values so asto increase the risk of true negative

88.83 95.33

10% variation from optimal values soas to increase the risk of true negative

87.50 95.33

• Locations and Sensors



______________________________________________________________________________________

(a) Main screen

(b) Threshold settings

(c) Emergency contact settings

Fig. 3: User interfaces in fall detection system.

Fig. 4: The content of a warning massage.

Due to smart phone internally installs multiple sensors,fall detection algorithm uses only acceleration and magne-tometer. Firstly, the experiment evaluates the accuracy ofacceleration in the chest and waist. From Fig. 5, it canbe seen that there is the better accuracy in the chest withaccuracy of 90.4%. Obviously, the movement of chest isthe better indication of fall detection. Secondly, when smartphone is placed in the waist, the accuracy with one sensoris lower than algorithm with two sensors. So, in this paper,both acceleration and magnetometer are used to detect thefall event.

• Performance ComparisonBy a series of experimental tests, the experimental results

are shown in Table III below. Table III shows the number ofsamples each activity in parenthesis. It can be observed thatthe accuracy of this algorithm is quite high as shown in TableIII. Among experimental data, walking, going down-stairs,and going upstairs are correctly identified and recognitionrate achieve the 100% in our data set. There are five

Fig. 5: Accuracy variance of different location and different numbers ofsensors.

jumping events mistaken for fall and two running eventsmistaken for fall. The main reason of the phenomenon is thatyoung subjects are different from older people in simulatingjumping and running activities. Finally, our algorithm canrecognize 109 falls from 120 simulated fall events.

TABLE III: The result of the experiment tests

SamplesWalking(30)

Goingdown-stairs(30)

Goingup-stairs(30)

Jumping(30)

Running(30)

Fall(120)

Fall 0 0 0 5 2 109Non-Fall 30 30 30 25 28 11

Both Cao’s method [35] and Fang’s method [36] arecompared to our method. Table IV lists the result of thethree methods and shows the best part in our algorithm’sresult. Cao’s method is based on three axis accelerations and



______________________________________________________________________________________

only considers the resultant acceleration values. The reasonthat we compare with Cao’s method is to illustrate that twosensors can obtain more information to distinguish the fallevent. As for the Fang’s method, it also puts the device inthe waist and we can compare the sensitivity and specificitywhen two devices are put in the similar position in aspectof performance. By contrast, we find that our algorithm hasbetter performance.

TABLE IV: The result of the three methods

Ours Cao Fang

Sensitivity (%) 90.83 92.75 72.22

Specificity (%) 95.33 86.75 73.78

D. Power Consumption

To test power consumption, we fully charged the twoSONY Z3 phones and then monitored the power statescontinuously for 6 hours respectively. One of them wasremained without the fall detection application, while theother ran the application continuously. Fig. 6 presents thetwo curves of battery level states versus time during the timeperiod of 6 hours. If the application keeps running normallyuntil the battery power is exhausted, it will last more than20 hours.

Fig. 6: Power consumption: the solid line shows the battery levels whenthe phone runs without the fall detection application; the dash line

presents the battery levels when the monitoring daemon and the patternrecognition in the fall detection system run.

VI. CONCLUSION

This paper proposed a fall detection algorithm that utilizesthe smart phone with built-in accelerometer and magne-tometer sensor to recognize the fall event from the humanactivities. We firstly select ten subjects simulate ADL andfall events to collect raw data which is to observe thedifference in human activities and get the option thresholds.Considering the trade-off between recognition accuracy andcomputing load, the sampling rate of 15Hz is adopted bythe algorithm. In addition, the algorithm uses the peak valueof SVM to detect fall events totally but there are still someADL mistaken for fall. In order to improve the accuracy,base length of the triangle (B) and post-impact velocity (V)are made the distinction between a real fall and some ADL,like jumping, quickly sitting on a bed or physical exercise.However, when the subject is running, the waveform ofwhich is similar to the waveform of the fall event meets thethresholds above. So, our approach considers the residual

movement to distinguish with the real fall event and runningactivity. Finally, the algorithm uses the vertical accelerationto monitor the subject’s body posture for verification.

REFERENCES

[1] W. H. O. Ageing and L. C. Unit, WHO global report on fallsprevention in older age. World Health Organization, 2008.

[2] S. Yoshida-Intern, “A global report on falls prevention epidemiologyof falls,” WHO. Geneva, 2007.

[3] R. Luque, E. Casilari, M.-J. Moron, and G. Redondo, “Comparisonand characterization of android-based fall detection systems,” Sensors,vol. 14, no. 10, pp. 18543–18574, 2014.

[4] M. Alwan, P. J. Rajendran, S. Kell, D. Mack, S. Dalal, M. Wolfe,and R. Felder, “A smart and passive floor-vibration based fall detectorfor elderly,” in Information and Communication Technologies, 2006.ICTTA’06. 2nd, vol. 1, pp. 1003–1007, IEEE, 2006.

[5] H.-W. Tzeng, M.-Y. Chen, and M.-Y. Chen, “Design of fall detectionsystem with floor pressure and infrared image,” in System Science andEngineering (ICSSE), 2010 International Conference on, pp. 131–135,IEEE, 2010.

[6] E. Vlaeyen, M. Deschodt, G. Debard, E. Dejaeger, S. Boonen, T. Goe-deme, B. Vanrumste, and K. Milisen, “Fall incidents unraveled: a seriesof 26 video-based real-life fall events in three frail older persons,”BMC geriatrics, vol. 13, no. 1, p. 103, 2013.

[7] L. Alhimale, H. Zedan, and A. Al-Bayatti, “The implementation ofan intelligent and video-based fall detection system using a neuralnetwork,” Applied Soft Computing, vol. 18, pp. 59–69, 2014.

[8] J.-L. Chua, Y. C. Chang, and W. K. Lim, “A simple vision-based falldetection technique for indoor video surveillance,” Signal, Image andVideo Processing, vol. 9, no. 3, pp. 623–633, 2013.

[9] R. Cucchiara, A. Prati, and R. Vezzani, “A multi-camera vision systemfor fall detection and alarm generation,” Expert Systems, vol. 24, no. 5,pp. 334–345, 2007.

[10] D. Anderson, J. M. Keller, M. Skubic, X. Chen, and Z. He, “Recog-nizing falls from silhouettes,” in Engineering in Medicine and BiologySociety, 2006. EMBS’06. 28th Annual International Conference of theIEEE, pp. 6388–6391, IEEE, 2006.

[11] W.-Y. Shieh and J.-C. Huang, “Speedup the multi-camera video-surveillance system for elder falling detection,” in Embedded Softwareand Systems, 2009. ICESS’09. International Conference on, pp. 350–355, IEEE, 2009.

[12] C.-W. Lin and Z.-H. Ling, “Automatic fall incident detection in com-pressed video for intelligent homecare,” in Computer Communicationsand Networks, 2007. ICCCN 2007. Proceedings of 16th InternationalConference on, pp. 1172–1177, IEEE, 2007.

[13] Z. Pan, Y. Zhang, and S. Kwong, “Efficient motion and disparityestimation optimization for low complexity multiview video coding,”

[14] B. Chen, H. Shu, G. Coatrieux, G. Chen, X. Sun, and J. L. Coatrieux,“Color image analysis by quaternion-type moments,” Journal of Math-ematical Imaging and Vision, vol. 51, no. 1, pp. 124–144, 2015.

[15] F. Nie and P. Zhang, “Fuzzy partition and correlation for image seg-mentation with differential evolution,” IAENG International Journalof Computer Science, vol. 40, no. 3, pp. 164–172, 2013.

[16] Y.-S. Huang and Y.-J. Wang, “A hierarchical temporal memory basedhand posture recognition method,” IAENG International Journal ofComputer Science, vol. 40, no. 2, pp. 87–93, 2013.

[17] M. F. Hashmi, A. R. Hambarde, and A. G. Keskar, “Robust image au-thentication based on hmm and svm classifiers,” Engineering Letters,vol. 22, no. 4, pp. 183–193, 2014.

[18] M. H. Taieb, J.-Y. Chouinard, and D. Wang, “Low complexity rateestimators for low latency wyner-ziv video decoders,” EngineeringLetters, vol. 21, no. 1, pp. 1–9, 2013.

[19] R. Perez-Aguila, “Enhancing brain tissue segmentation and imageclassification via 1d kohonen networks and discrete compactness: anexperimental study,” Engineering Letters, vol. 21, no. 4, pp. 171–180,2013.

[20] X. Yang, A. Dinh, and L. Chen, “Implementation of a wearerable real-time system for physical activity recognition based on naive bayesclassifier,” in Bioinformatics and Biomedical Technology (ICBBT),2010 International Conference on, pp. 101–105, IEEE, 2010.

[21] T. Zhang, J. Wang, L. Xu, and P. Liu, “Fall detection by wearablesensor and one-class svm algorithm,” in Intelligent Computing inSignal Processing and Pattern Recognition, pp. 858–863, Springer,2006.

[22] H. Gjoreski, S. Kozina, M. Gams, and M. Lustrek, “Rarefallreal-timeactivity recognition and fall detection system,” in Pervasive Computingand Communications Workshops (PERCOM Workshops), 2014 IEEEInternational Conference on, pp. 145–147, IEEE, 2014.



______________________________________________________________________________________

[23] P. T. Phu, N. T. Hai, and N. T. Tam, “A threshold algorithm in a fallalert system for elderly people,” in 5th International Conference onBiomedical Engineering in Vietnam, pp. 347–350, Springer, 2015.

[24] C. Wang, M. R. Narayanan, S. R. Lord, S. J. Redmond, and N. H.Lovell, “A low-power fall detection algorithm based on triaxial ac-celeration and barometric pressure,” in Engineering in Medicine andBiology Society (EMBC), 2014 36th Annual International Conferenceof the IEEE, pp. 570–573, IEEE, 2014.

[25] R. Tiwari, A. K. Singh, and S. N. Khan, “Using android platformto detect free fall,” in Information Systems and Computer Networks(ISCON), 2013 International Conference on, pp. 161–163, IEEE, 2013.

[26] W. Wibisono, D. N. Arifin, B. A. Pratomo, T. Ahmad, and R. M.Ijtihadie, “Falls detection and notification system using tri-axial ac-celerometer and gyroscope sensors of a smartphone,” in Technologiesand Applications of Artificial Intelligence (TAAI), 2013 Conference on,pp. 382–385, IEEE, 2013.

[27] Y. Kim, S. Kim, D. Kang, H. Park, N. Kim, S. H. Yang, and Y. Kim,“A simple falling recognition scheme for a human body by usingmobile devices,” in Proceedings of the 1st International Conferenceon Advanced Information and Comuter Technology (AICT), Warsawa,Poland, pp. 14–17, 2013.

[28] J. Dai, X. Bai, Z. Yang, Z. Shen, and D. Xuan, “Perfalld: A pervasivefall detection system using mobile phones,” in Pervasive Computingand Communications Workshops (PERCOM Workshops), 2010 8thIEEE International Conference on, pp. 292–297, IEEE, 2010.

[29] M. V. Albert, K. Kording, M. Herrmann, and A. Jayaraman, “Fallclassification by machine learning using mobile phones,” PloS one,vol. 7, no. 5, p. e36556, 2012.

[30] F. Sposaro and G. Tyson, “ifall: an android application for fallmonitoring and response,” in Engineering in Medicine and BiologySociety, 2009. EMBC 2009. Annual International Conference of theIEEE, pp. 6119–6122, IEEE, 2009.

[31] M. Dumitrache and S. Pasca, “Fall detection algorithm based on triax-ial accelerometer data,” in E-Health and Bioengineering Conference(EHB), 2013, pp. 1–4, IEEE, 2013.

[32] P. Kostopoulos, T. Nunes, K. Salvi, M. Deriaz, and J. Torrent,“Increased fall detection accuracy in an accelerometer-based algorithmconsidering residual movement,”

[33] J. Dai, X. Bai, Z. Yang, Z. Shen, and D. Xuan, “Mobile phone-basedpervasive fall detection,” Personal and ubiquitous computing, vol. 14,no. 7, pp. 633–643, 2010.

[34] M. Kangas, A. Konttila, P. Lindgren, I. Winblad, and T. Jamsa,“Comparison of low-complexity fall detection algorithms for bodyattached accelerometers,” Gait & posture, vol. 28, no. 2, pp. 285–291,2008.

[35] Y. Cao, Y. Yang, and W. Liu, “E-falld: A fall detection system usingandroid-based smartphon,” in Fuzzy Systems and Knowledge Discovery(FSKD), 2012 9th International Conference on, pp. 1509–1513, IEEE,2012.

[36] S.-H. Fang, Y.-C. Liang, and K.-M. Chiu, “Developing a mobilephone-based fall detection system on android platform,” in Computing,Communications and Applications Conference (ComComAp), 2012,pp. 143–146, IEEE, 2012.



______________________________________________________________________________________

fall detection algorithm based on triaxial accelerometer ... · fall detection algorithm based on...

Documents